Agent-flame 🔥

Agent Flame is a real-time visualization and coordination system for a decentralized network of autonomous agents operating in a peer-to-peer (P2P) mesh. Instead of relying on a central server or controller, the system is designed so that each “agent” (node) can independently communicate, share data, and adapt to the behavior of other nodes in the network. This creates a self-organizing ecosystem where intelligence and decision-making are distributed across the entire mesh.

The “Network Map” represents the live topology of this system—showing how agents are connected, how many nodes are online, and how data flows between them. Metrics like total agents, active nodes, connections, and message volume provide insight into network health and activity. A “nominal” status indicates that the system is stable, with no detected failures or disruptions.

The “Agent Roster” likely tracks individual agents, including their roles, states, or capabilities, while the “Global Log” captures a timeline of interactions and events across the network. Together, these components allow users to monitor emergent behavior, debug interactions, and understand how decentralized coordination unfolds in real time.

Overall, Agent Flame functions as both an orchestration layer and an observability tool for autonomous systems—useful for simulations, AI swarms, distributed computing experiments, or next-generation decentralized applications where resilience, scalability, and adaptability are critical.

Agent Flame was built as a lightweight, real-time system that combines a decentralized messaging layer with a reactive visualization frontend. At its core, each “agent” runs as an independent process (Node.js services) that communicates over WebSockets in a peer-to-peer style topology. Instead of routing everything through a single backend, agents broadcast state updates and messages to connected peers, forming a self-healing mesh where nodes can join/leave dynamically without breaking the network.

For transport and synchronization, we used a hybrid approach: WebSockets for low-latency streaming, plus a minimal event protocol (JSON-based) to standardize how agents announce presence, exchange messages, and update their state. A small coordination layer keeps track of active peers and connections, but avoids becoming a bottleneck by offloading most logic to the agents themselves.

The frontend is built with React and a real-time rendering layer (Canvas/WebGL via libraries like D3 or Three.js) to visualize the network graph. Nodes and edges are continuously updated based on incoming events, allowing us to display live connection counts, message flow, and system health. State management is event-driven, so the UI reacts instantly as agents communicate.

One particularly hacky—but effective—solution was simulating peer discovery using a lightweight relay/bootstrap node. New agents briefly connect to it to find peers, then switch to direct connections, keeping the system mostly decentralized while avoiding complex discovery protocols.